ORCA-CLEAN: A Deep Denoising Toolkit for Killer Whale Communication

ORCA-CLEAN, designed by Christian Bergler, is a deep denoising network designed for denoising of killer whale (Orcinus Orca) underwater recordings, not requiring any clean ground-truth samples, in order to improve the interpretation and analysis of bioacoustic signals by biologists and various machine learning algorithms.
ORCA-CLEAN was trained exclusively on killer whale signals resulting in a significant signal enhancement. To show and prove the transferability, robustness and generalization of ORCA-CLEAN even more, a deep denoising was also conducted for bird sounds (Myiopsitta monachus) and human speech.

As described in our manuscript, entitled ORCA-CLEAN: A Deep Denoising Toolkit for Killer Whale Communication we decompressed the 256 x T frequency-compressed network output to a 2,049 x T denoised spectrogram, which we used as a spectral mask, multiplied it with the original 2,049 x T complex spectrum, and transformed the spectral content back to the audio domain.

For each domain - birds, human speech, and orcas - we provide the spectral content as well as the denoised audio files to show and prove the proposed auditive and visual cross-domain ability of ORCA-CLEAN while not requiring any clean ground-truth labels.

In the following, the original audio files and the denoised results of ORCA-CLEAN are visualised as spectrograms (FFT-size = 4.096, Hop-size = 441, SamlingRate = 44.1 kHz). For each spectrogram the corresponding audio files were stored for listening. The audio files can be listened to by clicking on the spectrograms.

All denoised examples listed in the manuscript (Figure 5 and Figure 6) are part of the listed multimedia examples. For each manuscript example, the exact location within the multimedia folder is mentioned.

Killer Whale: (Orcinus Orca):

11 different noise-heavy underwater sound examples of orcas (11 vocalization categories including 1 example per category) within a frequency range of 800 Hz up to 10 kHz (see Figure 5 manuscript and description).

Bird: (Myiopsitta monachus):

4 different noise-heavy sound examples of bird sounds within a frequency range of 1 kHz up to 8 kHz (see Figure 6 manuscript).

Human Speech:

6 different noise-heavy sound examples of 2 human speakers (3 female, 3 male) within a frequency range of 0 kHz up to 4 kHz (see Figure 6 manuscript).

Paper Preprint:

Attached please find a PREPRINT VERSION of our submitted manuscript, entitled "ORCA-CLEAN: A Deep Denoising Toolkit for Killer Whale Communication", submitted to InterSpeech 2020.

ORCA-CLEAN PREPRINT